QCEFF RTPS inflate

[hkershaw-brown]

In filter_ensemble_inflate inflation does transform to/from probit.

non-RTPS inflation:

transform_to_probit
inflate
transform_from_probit

DART/assimilation_code/modules/assimilation/filter_mod.f90

Lines 1594 to 1616 in 464aa57

	! This is an initial test of doing inflation in probit space
	! Note that this appears to work with adaptive inflation, but more research would be good
	! Probably also shouldn't be used with groups for now although it is coded to do so
	do j = 1, ens_handle%my_num_vars
	call get_state_meta_data(ens_handle%my_vars(j), my_state_loc, my_state_kind)
	! Need to specify what kind of prior to use for each
	call probit_dist_info(my_state_kind, .true., .true., dist_type, &
	bounded_below, bounded_above, lower_bound, upper_bound)
	call transform_to_probit(grp_size, ens_handle%copies(grp_bot:grp_top, j), &
	dist_type, dist_params, probit_ens(1:grp_size), .false., &
	bounded_below, bounded_above, lower_bound, upper_bound)
	! Compute the ensemble mean in transformed space
	probit_ens_mean = sum(probit_ens(1:grp_size)) / grp_size
	! Inflate in probit space
	call inflate_ens(inflate, probit_ens(1:grp_size), probit_ens_mean, &
	ens_handle%copies(inflate_copy, j))
	! Transform back from probit space
	call transform_from_probit(grp_size, probit_ens(1:grp_size), &
	dist_params, ens_handle%copies(grp_bot:grp_top, j))
	end do

RTPS inflation:

inflate

DART/assimilation_code/modules/assimilation/filter_mod.f90

Lines 1575 to 1592 in 464aa57

	if ( do_rtps_inflate(inflate)) then
	if ( present(SPARE_PRIOR_SPREAD) .and. present(ENS_SD_COPY)) then
	write(msgstring, *) ' doing RTPS inflation'
	call error_handler(E_MSG,'filter_ensemble_inflate:',msgstring,source)
	!Reset the RTPS factor to the given input.nml value
	ens_handle%copies(inflate_copy, 1:ens_handle%my_num_vars) = inf_initial(POSTERIOR_INF)
	do j = 1, ens_handle%my_num_vars
	call inflate_ens(inflate, ens_handle%copies(grp_bot:grp_top, j), &
	ens_handle%copies(ENS_MEAN_COPY, j), ens_handle%copies(inflate_copy, j), 0.0_r8, &
	ens_handle%copies(SPARE_PRIOR_SPREAD, j), ens_handle%copies(ENS_SD_COPY, j))
	end do
	else
	write(msgstring, *) 'internal error: missing arguments for RTPS inflation, should not happen'
	call error_handler(E_ERR,'filter_ensemble_inflate',msgstring,source)
	endif
	else

[hkershaw-brown]

Need to add RTPS to tests.
Unclear if it is intentionally not being probit transformed (incompatible?) or an oversight.

[hkershaw-brown]

notes from conversation: RTPS is not compatible with QCEFF.

[jlaucar]

This issue was discussed in the code review of the QCEFF inflation issue by Moha yesterday. For now, we do not understand whether RTPS could be implemented in a transformed space or not. The code in the QCEFF patch will continue to do RTPS in the same way, whether a probit transform is selected for the inflation or not. Should this be documented somewhere besides in the code?

…

On Fri, Dec 20, 2024 at 7:30 AM Helen Kershaw ***@***.***> wrote: notes from conversation: RTPS is not compatible with QCEFF. — Reply to this email directly, view it on GitHub <#748 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANDHUISR2LI3SM4HUCVYQGT2GQSYPAVCNFSM6AAAAABPJZ2P42VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNJXGEZDGMRRGA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[hkershaw-brown]

Documented in the documentation and code corrected, right? The code is silently ignoring QCEFF at the moment. If someone selects RTPS inflation, should qceff be disabled?

I'm assuming the QCEFF patch is this branch: https://github.com/NCAR/DART/tree/qceff_inflation_fix ? So separate issue to RTPS?

[jlaucar]

This should be a separate issue. The RTPS can be used any observation space distribution and with the probit transform for both the state and observed variables during the assimilation. It is only the inflation specification that is a no-op. We are going to need separate issues for this question and for the group filter in combination with the QCEFF. Lili has used this but we have never tested it and there are several places where what we're doing in the case of failed transposes is questionable. Where would this be best documented?

…

On Fri, Dec 20, 2024 at 8:30 AM Helen Kershaw ***@***.***> wrote: Documented in the documentation and code corrected, right? The code is silently ignoring QCEFF at the moment. If someone selects RTPS inflation, should qceff be disabled? I'm assuming the QCEFF patch is this branch: https://github.com/NCAR/DART/tree/qceff_inflation_fix ? So separate issue to RTPS? — Reply to this email directly, view it on GitHub <#748 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANDHUIVDGPCDWBOBTDGCC3D2GQZ27AVCNFSM6AAAAABPJZ2P42VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNJXGIZDKMZRGM> . You are receiving this because you commented.Message ID: ***@***.***>

[hkershaw-brown]

Yup, this RTPS issue is its own issue (#748) separate from the other QCEFF issues
https://github.com/NCAR/DART/issues?q=is%3Aopen+is%3Aissue+label%3AQCEFF

For the group filter combo with QCEFF - the best place to document an issue is to make a new issue in github. Label it QCEFF so someone can look at all the QCEFF issues if needed, and we can close the separate issues as they get fixed.

[jlaucar]

Just a quick note that RTPS is spelled at RPTS in issue #748. Only important because a search of issues for "RTPS" returns nothing. Not sure if it is trivial, worthwhile, or even possible, to switch that to RTPS.

…

On Fri, Dec 20, 2024 at 9:01 AM Helen Kershaw ***@***.***> wrote: Yup, this RTPS issue is its own issue (#748 <#748>) separate from the other QCEFF issues https://github.com/NCAR/DART/issues?q=is%3Aopen+is%3Aissue+label%3AQCEFF For the group filter combo with QCEFF - the best place to document an issue is to make a new issue in github. Label it QCEFF so someone can look at all the QCEFF issues if needed, and we can close the separate issues as they get fixed. — Reply to this email directly, view it on GitHub <#748 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANDHUIUX6GBC7W3RZW4SY7T2GQ5M7AVCNFSM6AAAAABPJZ2P42VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNJXGI3TMMJUHA> . You are receiving this because you commented.Message ID: ***@***.***>

[hkershaw-brown]

thanks @jlaucar they look exactly the same to me, I'll change the title.